Document not found! Please try again

automated brain hemorrhage lesion segmentation and classification ...

1 downloads 0 Views 2MB Size Report
Aug 10, 2015 - Accurate identification of acute hemorrhage is a critical step in planning appropriate therapy. The correct characterization of underlying ...
Journal of Theoretical and Applied Information Technology 10th August 2015. Vol.78. No.1

© 2005 - 2015 JATIT & LLS. All rights reserved.

ISSN: 1992-8645

www.jatit.org

E-ISSN: 1817-3195

AUTOMATED BRAIN HEMORRHAGE LESION SEGMENTATION AND CLASSIFICATION FROM MR IMAGE USING AN INNOVATIVE COMPOSITE METHOD 1

SUDIPTA ROY, 2SANJAY NAG, 3SAMIR KUMAR BANDYOPADHYAY, 4

1, 2,3

DEBNATH BHATTACHARYYA, 5TAI-HOON KIM Department of Computer Science and Engineering, 4Department of Computer Application, 5

Department of Convergence Security

1

2,3

Academy of Technology, Adisaptagram, Hooghly-712121, West Bengal, India Calcutta University Technology Campus, JD-2, Sector-III, Salt Lake, Kolkata-98, India 4

5

Vignan's Institute of Information Technology, Visakhapatnam-530049, AP, India Sungshin Women's University, 249-1, Dongseon-dong 3-ga, Seoul, 136-742, Korea

E-mail: [email protected], [email protected], [email protected], [email protected], 5

[email protected]

ABSTRACT Accurate identification of acute hemorrhage is a critical step in planning appropriate therapy. The correct characterization of underlying pathology, such as neoplasia, vascular malformation, or infarction, is equally important for conclusive diagnosis. Computer-aided diagnosis (CAD) systems have been the focus of several research endeavors and it is solely based on the idea of analyzing images of different types of brain hemorrhage by implementing improved image processing algorithms. In this paper, an innovative and automated approach is cited to detect existence of brain hemorrhage in multiple MRI scan images of the brain, the type of hemorrhage and position of the hemorrhage. The implemented algorithm includes several stages such as artifact and skull elimination constituting preprocessing, image segmentation, and hemorrhage localization. The result of the conducted experiments has been compared statistically with reference image showing very promising results. Keywords: Brain, Hemorrhage, MRI, CAD, Hemorrhage Lesion, Image Segmentation, Position of Hemorrhage, Classification 1.

depicts anatomy in far greater details while at the same time being more sensitive and specific to abnormalities within the brain. Medical image segmentation attempts to label pixels by tissue type. Image segmentation play a significant role in many medical applications ranging from the education, assessment of medical students to image-guided surgery and surgical simulation. Appropriate segmentation methods have high correlation with image acquisition modality, the tissue of interest and since MRI provides superior contrast of soft tissue structures it is the choice of method for imaging the brain. Most research work on brain segmentation focuses on MRI for the same reason. The detection of hemorrhage is essential in the diagnosis and management of a variety of intracranial diseases including hypertensive hemorrhage, hemorrhagic infarction, brain tumor,

INTRODUCTION

In contemporary times, cerebrovascular diseases are the third cause of death in the world after cancer and heart diseases [1]. A common cerebrovascular diseases is brain hemorrhage which is caused by busting of one or more blood vessels within brain causing bleeding. There are many types of brain hemorrhage such as: epidural, subdural, subarachnoid, cerebral, and intra-parenchymal hemorrhage. The different types stated differ in many aspects such as the size of the hemorrhage region, its shape, and its location. MRI and CT are equally helpful for determining when hemorrhage is present [2]. MRI is more sensitive than CT for differentiating other types of neurological disorders that mimic the characteristics of stroke. MRI has a significant range of available soft tissue contrast, 34

Journal of Theoretical and Applied Information Technology 10th August 2015. Vol.78. No.1

© 2005 - 2015 JATIT & LLS. All rights reserved.

ISSN: 1992-8645

www.jatit.org

E-ISSN: 1817-3195

and robustly. The segmentation of light abnormality within the high background intensity performs unsatisfactorily result. It is observed that segmentation of dark abnormalities is not as effective as segmentation in light abnormalities. [10] Generalized fuzzy c-means algorithm uses both pixel attributes and spatial local information that is weighted in correspondence with neighbor elements based on their distance attributes. This has the potentiality to improve the segmentation performance tremendously. This improves the segmentation performance dramatically. Poor contrast, noise and non-uniform intensity variation can affect the results. The important phases of [11] paper are feature extraction, reduction of dimensionality, unsupervised data clustering, voxel classification and interactive post processing refinement. A result of the unsupervised 3D segmentation seems to be highly stable but a comparison with standard unsupervised methods (k-means) is not very significant in the clinical environment as a consequence of the segmentation of multivariate medical images. The color converted segmentation with Kmeans clustering algorithm and regions of the brain related to hemorrhage can be correctly separated from colored image and it help pathologist to distinguish lesion size and its region exactly. A first detection process [12] is based on selecting asymmetric areas with respect to the approximate brain symmetry plane. Its application to several datasets with different abnormality sizes, intensities and locations shows that it can automatically detect and segment very different types of brain abnormality with a good quality. A symmetric based [13] result constitutes the initialization of a segmentation method based on a combination of a deformable model and spatial relations, leading to a precise segmentation of the abnormalities. In the last phase of their work, the hemorrhage regions were determined according to their locations and gray level statistics. The result obtained is extremely encouraging; however, the authors acknowledged that it is a premature work. An enhancement process is applied for improving the quality of images along with mathematical morphology to increase the contrast in MRI images. Wavelet transform is applied in segmentation process for decomposition of MRI images [14]. The algorithms proposed in this research can contribute towards education of senior medical students/ resident doctors/ radiology students and further research in this field. The practical utility however is that it can assist in detecting hemorrhage through large sets of MRI of brain scans within very small time period.

cerebral aneurysm, vascular malformation, trauma, hemorrhagic changes following radio- or chemotherapy, and hemorrhagic pial metastasis. A recent report has indicated that the detection of hemorrhage on MR images is useful for the grading of gliomas. It has been found that T2-weighted gradient-recalled echo sequence is more sensitive than the T2-weighted spin-echo and fast SE sequences. The eason for this is the magnetic susceptibility induced by static field inhomogeneities arising from paramagnetic blood breakdown products. CAD systems incorporate computers to add a new dimension to physicians for achieving a faster and more accurate diagnosis. CAD systems are usually domain-specific as they are optimized for certain types of diseases, focuses on specific parts of the body and diverse diagnosis methods. They analyze different kinds of input such as stated symptoms, laboratory tests results, medical images, etc. corresponding to their specific domain. Diagnoses involving medical image interpretation systems are very useful since they can be integrated with the associated firmware and software accompanying the medical imaging machine to provide efficient and accurate diagnosis. Development of such CAD system is challenging since they combine the elements of artificial intelligence and digital images processing. This work proposes a CAD system to assist the radiologist for the detection of hemorrhages in MRI scan images of human brain and identify their types. Some of the older works [3, 4] addressed the problem of segmenting the region of intracerebral hemorrhages. The former work used a spatially weighted k-means histogram-based clustering algorithm, whereas in the latter work, the authors applied a multi resolution simulated annealing method. Cheng and Cheng [5] proposed a Fuzzy CMeans (FCM) method based on multi resolution pyramid for brain hemorrhage analysis. They also compared FCM, competitive Hopfield neural network [6] and fuzzy Hopfield neural network [7] in the global thresholding stage. In another work, Liu et al. [8] proposed an alternative fuzzy C-means method for the segmentation phase. A more contemporary work based on FCM is the work of Li et al. [9], used a thresholding technique based on FCM clustering to remove all non-brain regions. The results of this clustering were brain regions that were segmented into slices using median filtering to eliminate noise in the image, and then the maximum entropy threshold was computed for each slice to determine the potential hemorrhage regions. Region growing is not often used alone because it is not sufficient to segment brain structures accurately 35

Journal of Theoretical and Applied Information Technology 10th August 2015. Vol.78. No.1

© 2005 - 2015 JATIT & LLS. All rights reserved.

ISSN: 1992-8645

www.jatit.org

Standard deviation, denoted as σ, equals the square root of the variance s-squared is written below

This paper is organized as follows: In this section a brief overview of related works was discussed. The proposed methodology is discussed in section 2. In section 3, the experimental results are portrayed to assess the quality of the proposed method on test dataset. In section 4, performance evaluation and accuracy measurement with ground truth image has been presented and finally we conclude our paper in section 5. 2.

σ

The data set of “Whole Brain Atlas” image data base [15] which consists of T1 weighted, T2 weighted, proton density (PD) MRI image containing multiple image slice has been used. The RGB image has been converted to grayscale image using a weighted sum of the R, G, and B components multiplied by a constant. The transformation function is given below

f1 x, y

f x, y



1 if f x, y ! σ 0 if f x, y # σ

(6)

$ % 1 & (7) Reversing the intensity levels of an image in this manner produces the equivalent of a photographic negative and this type of processing is particularly suited for enhancing white or gray detail embedded in dark regions of an image, especially when the black areas are dominant in size. For binary image complement the algorithm use f2(x,y)=1-f1(x,y) which prepares it for the next step of wavelet decomposition. We begin by defining the wavelet series expansion of function f2(x) ϵ L2(R) relative to wavelet ψ(x) and scaling function φ(x). f2(x) can be represented by a scaling function expansion and some number of wavelet function expansions in sub-spaces W() , W() * , W()+, ….. Thus

s T r (2) For maintaining simplicity in notation, r and s are variables denoting, respectively, the gray level of f(x, y) and g(x, y) at any point (x, y). This conversion is followed by image binarization, constituting the preprocessing step and threshold intensity is calculated by standard deviation of the image pixel intensity. To calculate the standard deviation, mean and variance derivation is written below. Mean is defined as the division of the number of samples multiplied by the sum of all data points ∑

(5)

The negative of an image with gray levels in the range [0, L-1] is obtained by using the negative transformation is given by the expression

g x, y T f x, y (1) Here f(x, y) is the input image, g(x, y) is the processed image, and T is an operator on f, defined over some neighborhood of f(x, y). In addition, T can operate on a set of input images. The simplest form of T is when the neighborhood is of size 1×1 (that is, a single pixel). In this case, g depends only on the value of f at (x, y), and T becomes a graylevel (also called an intensity or mapping) transformation function of the form



√v

The obtained standard deviation intensity value is used as threshold intensity to binarize the MR Image of brain and is very much helpful for extracting brain portion and differentiating it from to the non-brain portion. MRI of brain has the significant intensity difference between background and the foreground so the use of standard deviation based binarization has been successfully implemented for brain stroke detection purpose.

PROPOSED METHODOLOGY

μ

E-ISSN: 1817-3195

f2 x

5

. c() k φ() x 3 . . d( k ψ(,2 x 2

( () 2

(8)

Where j0 is an arbitrary starting scale and the c() (k) and dj(k) are relabeled. The c() (k) normally called approximation or/and scaling coefficients; the dj(k) are referred to as detail or/and wavelet coefficients. Thus in the above equation first sum uses scaling function to provides an approximation of f2(x) at scale j0. For each higher scale j ≥ j0 in the second sum, a finer resolution function a sum of wavelet is added to the approximation to provide increasing details. If the expansion function forms an

(3)

Variance, is denoted as v, equals 1 divided by the number of samples minus one, multiplied by the sum of each data point subtracted by the mean then squared. ∑ ∑ v f x, y μ (4) 36

Journal of Theoretical and Applied Information Technology 10th August 2015. Vol.78. No.1

© 2005 - 2015 JATIT & LLS. All rights reserved.

ISSN: 1992-8645

www.jatit.org

i AH, V, DE (12) As in the 1-D case, j0 is an arbitrary starting scale and the WF j , m, n coefficients define an approximation f2(x, y) at scale j0. The WJK j, m, n coefficients add horizontal, vertical, and diagonal details for scales j≥j0. normally j0=0 and N=M=2J so that j=0,1,2,….,J-1 and m=n=0,1,2,….,2j-1. Thus f2(x, y) is obtained via the inverse discrete transform as obtained by the equation

orthogonal basis or tight frame, which is often the case, the expansion coefficients are calculated and is shown by the equations below c() k

7 f2 x , φ() x ! 8 f2 x φ() x dx

and d( k

E-ISSN: 1817-3195

7 f2 x , ψ(,2 x ! 8 f2 x ψ(,2 x dx (9)

Above two coefficients expansion are defined as inner products of function being expanded and the expansion functions being used where φ() and ψ(,2 are the expansion functions; c() and d( are the expansion coefficients. Two dimensional (2-D) scaling function, φ(x,y), which is a product of two 1-D functions and three two dimensional wavelets, ψH(x,y), ψV(x,y), and ψD(x,y) are required. Excluding the products that produce 1-D results, like φ(x) ψ(x), the four remaining products produce the separable scaling function and separable directionally sensitive wavelets φ x, y

. . WF j , m, n φ(),